An unsupervised approach to exploring speaking task complexity based on fluency metrics

Authors: Neda Mousavi, Sven Grawunder

Abstract:

In this study, task complexity is investigated by analyzing the fluency metrics extracted based on the temporal patterns of verbal and non-verbal elements in two languages. The data include a speech corpus of 60 participants (30 Persian and 30 German speakers) who completed seven different speaking tasks, including reading formal and informal texts, spontaneous conversations, describing pictures, telling stories, and leaving formal and informal messages. The recorded audio files were annotated using four labels - speech, pause, filler and repair - and 19 metrics for fluency were extracted based on duration, number, rate, and ratio of intervals. In the first step, principal component analysis revealed differences between the two languages in how tasks were distributed in the PCA space. This was followed by KMeans clustering, applied as an unsupervised method to identify hidden patterns, which were interpreted in relation to task complexity. The model identified four complexity clusters in both languages, with distinct distribution patterns. German speakers exhibited a more structured clustering, indicating greater adaptation to task demands, while Persian speakers showed a less regular distribution, suggesting weaker adherence to task genres and potentially greater individual variability.